How the engine decides

What gets read

Three kinds of source, all of them about the company rather than about the market. SEC filings, which is to say 10-K annual reports, 10-Q quarterly reports and 8-K event disclosures. Earnings call transcripts, where management discusses the quarter it has just reported. And material news, filtered hard.

News gets the most scepticism because most of a news feed is not information about a business. A publisher denylist removes the price-target mills and rank aggregators, a relevance floor removes articles that merely mention the ticker, and a triage pass asks one question of what survives: is this a development at the company, or is it commentary? No sentiment score is read or stored. A sentiment label is an opinion about the market, not evidence about a business.

Company data is fetched once and shared. If a hundred investors track the same company, one new filing is downloaded, parsed and summarised a single time. Only the judgement against each individual thesis runs a hundred times, because only that part differs.

What a thesis is judged against

Not everything ever published. When you write a thesis, Okidor finds the most recent 10-K or 10-Q published on or before that day, which is the set of accounts you were reading when you formed the view. That filing becomes the baseline, and everything published between it and the day you wrote is read as part of the same review rather than as a series of separate verdicts.

Nothing earlier is ever audited. What happened three sets of accounts ago is not evidence about a thesis written last week, and burying the one audit that matters under a wall of “no material impact” is worse than showing nothing at all.

Who assigns the verdict

Not the model. This is the part worth being precise about, because it is the difference between a tool you can rely on and a very fluent opinion.

The language model is given your thesis and the source, and it returns structured evidence: which of your stated assumptions the source bears on, which way, the passages it drew that from, and a directional reading. The output schema it is allowed to return has no verdict field in it at all.

The verdict is then computed by fixed rules, in a fixed order. A breached numeric criterion outranks everything. Then a breached qualitative criterion. Then a reading of weakening, then a criterion at warning, then a mixed reading. Only after all of those does anything positive get assigned. The report tells you which rule fired.

Two consequences follow. The same inputs always produce the same verdict. And a numeric criterion that has been breached cannot be talked out of it by prose, because the arithmetic happened before the model was asked anything.

How a sell rule is actually checked

It depends on what kind of rule it is, and the distinction matters.

A rule with a number in it, such as “gross margin drops below 40 percent”, is evaluated arithmetically against the figures in the filing itself. Companies publish their financials as XBRL alongside the readable document, so the margin is computed from the company’s own reported numbers. No model is asked whether it thinks the margin fell.

A rule that is a judgement, such as “a major customer moves to a competitor”, is assessed from the evidence in the document, and the audit quotes the passage it relied on so you can disagree with it.

The five verdicts

Every completed audit ends in exactly one of these, and they are ordered by how much they should worry you.

• Kill criterion triggered. A condition you wrote down as thesis-ending has been met.
• Watch signal. Evidence undermines an assumption without breaching a rule, or a criterion has moved to warning, or the source cuts both ways.
• Thesis holds. The results the thesis rests on are consistent with it. Only a periodic report can earn this, because it is a claim about the business rather than about an event.
• Supporting evidence. A single event that points your way. News and 8-K disclosures land here at most, however well they read.
• No thesis impact. Read in full, and it does not bear on anything you wrote. Most news is this, and that is the system working.

The gap between the last two is deliberate. A news article is one third party’s account of one event, and it cannot certify that a thesis holds. News also arrives far more often than filings, so without that cap a single favourable article would overwrite a watch signal earned by a quarterly report and tell you everything was fine.

Why most audits say nothing changed

Long-term investors do not need to react to every quarterly development. They need to know when a development actually affects the reasons they invested. The noise is overwhelming and the signal is rare, so separating the two is most of the work.

That is why “no thesis impact” is the commonest verdict the engine reaches, and why it is a sign of the system working rather than failing. A tool that found something urgent in every filing would be telling you about the filing, not about your thesis. You do not need more information. You need to know when something actually matters to what you wrote down.

News that turns out to bear on nothing does not become a row in your stream at all. It is counted, so the page can tell you how many items were read and that none of them mattered, and the reading stays available if you want it.

Writing the thesis in the first place

The quality of an audit depends on the precision of the thesis. A vague thesis produces vague judgements, and the part investors are laziest about is the part the engine leans on hardest: the conditions under which you would sell.

So there is a thought partner in the writing step, and it is deliberately not a drafting tool. Before its first reply it loads real grounding for the company you picked, then it argues with you: it challenges assumptions, asks what would have to be true, and pushes on the sell rules. It runs for three exchanges and then summarises what it understood, which you edit.

The reason it works this way is that a thesis the AI wrote is worthless. You cannot be held to reasoning you did not do. A thesis you wrote and then had to defend is the thing worth auditing.

What happens when it fails

An audit that cannot be completed is recorded as a failure and shown as one. It is never dressed up as a result, and the database will not accept a verdict on a run that did not finish, so a pipeline outage cannot surface as “thesis holds”.

This rule exists because the system once produced the string “Material event filed. Thesis evaluation pending full review” and presented it as an answer. A tool that tells you nothing is annoying. A tool that tells you nothing while looking like it told you something is dangerous.

What it does not do

It does not recommend buying or selling anything, it does not score companies, and it does not predict prices. It has no opinion about whether your thesis was a good one. It reports whether the things you said you were relying on are still true, and leaves the decision where it belongs.

See a real audit or read about writing kill criteria.